Index Company
DeepSeek
15 CÈ pieces since 9 May 2026: 10 intel items, 5 essays. Newest first. All companies →
-
The model was never the moat: DeepSeek open-sources its agent harness under MIT
DeepSeek published Harness v0.1 under the MIT licence, built on the Cordis meta-framework where models, tools, sandboxes, sessions, and orchestration are all plugins. It ships OS-level sandboxing and supports Anthropic, OpenAI, Bedrock, Azure, and Gemini.
-
DeepSeek’s API price quadruples Sunday. Peak hours are Singapore’s working day.
V4-Pro-0813 shipped on 13 August with a one-million-token context and an MIT-licensed agent harness. Three days later the promotional rate that put it into OCBC’s toolset and Indosat’s Indonesian stack expires.
-
The cheap model beats the flagship: DeepSeek's V4-Flash-0731 wins all nine agent benchmarks
DeepSeek published the retrained V4-Flash-0731 to Hugging Face on Friday and opened its API beta. The build scores 82.7 on Terminal Bench 2.1 against 72.1 for the larger V4-Pro-Preview, at $0.14 per million input tokens.
-
Anthropic prices what you sell at $15 a million tokens. You get no vote.
On the dichotomy of control, and the founder whose gross margin is reset each morning by a price she gets no vote in.
-
Tokens now have a rush hour: DeepSeek doubles V4 API rates during Beijing business hours
DeepSeek moved V4 to general availability on 20 July with peak pricing that doubles per-token rates from 9am to noon and 2pm to 6pm Beijing time, leaving off-peak rates unchanged.
-
Grok 4.5 cut coding-AI prices by roughly 80 percent. Where to point it first.
xAI, OpenAI, and Meta all cut inference prices in the same week, and the cheapest coding model is now roughly twice as willing to make things up as the one it replaced.
-
With Nvidia's China share collapsing and Huawei near 50%, DeepSeek designs its own AI chip
DeepSeek is designing an in-house AI chip to cut its reliance on Nvidia and Huawei silicon, joining Alibaba and Baidu as US export controls erode Nvidia's China position and lift Huawei toward half the market.
-
The biggest customer the AI labs have is quietly building their replacement
Microsoft began routing tens of thousands of Excel and Outlook prompts a week to its in-house MAI models this month. The question that leaves for anyone building on AI is not which lab to trust.
-
Chinese open models from DeepSeek and Z.ai reach US frontier parity
US enterprises are moving workloads onto open-weight models from DeepSeek and Z.ai, now rated competitive with leading American systems, as OpenAI and Anthropic inference prices keep climbing.
-
DeepSeek open-sources DSpark, lifting V4 inference speed up to 85 percent
Released under an MIT licence, the DeepSpec framework pairs a small draft model with the main model to verify tokens in batches, accelerating DeepSeek's production V4 models with no hardware upgrade or retraining.
-
Chinese open-weight models pass 60 percent of OpenRouter's token volume
By May 2026, openly licensed models from Alibaba, DeepSeek, Moonshot, and Zhipu accounted for roughly 61 percent of tokens consumed on OpenRouter, with four of the five most-used models built in China.
-
Your whole company runs on one input. A memo can delete it by morning.
On premeditatio malorum, and the founder whose single dependency a regulator can delete with one memo overnight.
-
Chinese AI models run at one-sixth US cost, gain SEA adoption via open weights
Chinese models including DeepSeek V3, Qwen 3.7 Max, and MiniMax 3 now match Western flagship performance at one-sixth to one-quarter the inference cost, according to a RAND report published early 2026, and are spreading across Southeast Asia through open-weight releases and regional data-center investment.
-
MiniMax M3 ships one-million-token coding model but withholds training code
Chinese startup MiniMax launched M3 on June 1, 2026 with claims of topping GPT-5.5 and Gemini 3.1 Pro on benchmarks, yet has not released training code or inference operators, leaving it in open-weight territory rather than fully open source.
-
Four Chinese labs ship frontier-grade open-weights coding models in 12-day window
Z.ai's GLM-5.1, MiniMax M2.7, Moonshot's Kimi K2.6 and DeepSeek V4 all landed within a fortnight, hitting near-frontier benchmarks on agentic coding at under a third of Claude Opus 4.7's inference cost.